JPH1021398A - Method for extracting directional characteristic vector - Google Patents

Abstract

PROBLEM TO BE SOLVED: To exactly identify a character by exactly extracting the directional characteristic vector of an outline in an oblique line. SOLUTION: An outline is extracted as each horizontal, vertical, right-up, and right-down directional pattern from black picture elements constituting the outline picture of a character to be recognized, the outline picture is normalized, and each above mentioned directional pattern is processed based on this normalized information, and generated as each character stroke. Then, the generated character stroke of each direction is equally divided into small areas by each horizontal and vertical dividing line, and at the time of extracting counting the number of the black picture elements of each divided small area, and extracting it as a directional characteristic vector, the picture of the generated character stroke of each right-up or right-down direction is rotated, matched with the direction of the dividing line, and the character stroke after rotation is divided into small areas.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition technique for recognizing a character to be identified such as a handwritten character, and more particularly, to extracting each direction pattern from the outline of the character to be identified and normalizing the outline. The present invention relates to a directional feature vector extraction method for extracting a directional feature vector indicating a directional feature of a pattern from each directional pattern of a contour processed based on the normalization information.

[0002]

2. Description of the Related Art In general, character recognition is performed according to a procedure in accordance with a recognition processing flow shown in FIG. That is, first, in a preprocessing step S1, a noise component is removed from an original image of a character to be recognized. Next, in a direction pattern creation step S2, a contour image is obtained from the original image of the recognized character from which the noise component has been removed, and the contour image is divided into pixels, and a black pixel having character information or character information is stored. It is configured as a binary pixel of a white pixel which is not used.
Then, the direction pattern of the contour image is extracted from the binary image including the black pixels or the white pixels.

After the direction pattern is extracted in this way,
In the next normalization step S3, each directional pattern of the image is scaled up or down at a fixed ratio so that the image from which the directional pattern is extracted fits within a square frame of a predetermined size. In the next feature vector creation step S4, each direction pattern of the normalized image is equally divided into 8 × 8 small areas (square areas), and the number of black pixels in each small area is used as a feature amount. Then, a feature vector of each direction pattern is created.

[0004] Next, in a large classification step S5, a candidate character having a vector having a short distance among vectors in a dictionary provided in advance is compared with the directional feature vector of the input character (character to be recognized) obtained in step S4. A plurality is extracted in order from the closest one. This distance calculation is performed using a general calculation method called 4-neighbor-distance shown in Expression (1).

That is, d4 ((i, j), (h, k)) = | i−h | + | j−k | (1) (where (i, j) is the coordinate value of the input character, (h, k) are the coordinate values of the characters in the dictionary.) Then, in the detailed classification step S6, a detailed comparison is performed using a fine pattern (for example, a separate 32 × 32-dimensional feature vector is prepared) to make the final comparison. Perform character identification.

[0006]

Conventionally, when a directional feature vector is extracted, the horizontal and vertical lines of the normalized contour are used.
For each of the vertical, upward-sloping, and downward-sloping patterns, the pattern is equally divided into 8 × 8 square areas, and the number of pixels of each equally divided small area is counted. Directional feature vector. For this reason, continuous information can be obtained for each character stroke of a contour line having a horizontal or vertical component that matches the direction of the dividing line that is equally divided into each of the above-described small areas. Directional feature vectors.

However, in the case of a character stroke having a component in an oblique direction such as rising to the right or falling to the right, as shown in FIG.
To e9 do not match the direction of the dividing line,
Stroke information is dispersed in adjacent small areas and cannot be obtained as continuous stroke information. For this reason, there was a problem that an accurate directional feature vector could not be extracted, and thus accurate character identification could not be performed. Therefore, an object of the present invention is to perform accurate character identification by accurately extracting a direction feature vector of an oblique outline.

[0008]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention provides a method for forming a contour line from black pixels constituting a contour image of a character to be identified in horizontal, vertical, upward and downward directions and downward and upward directions. In addition to extracting as a pattern, normalizing the contour image and processing each of the above directional patterns based on the normalized information to generate each character stroke, the generated character stroke in each direction is divided into horizontal and vertical divisions. In the case where lines are equally divided into small regions, and the number of black pixels in each divided small region is counted and extracted as a directional feature vector, the generated character stroke image is made to coincide with the direction of the dividing line. And the character stroke after rotation is equally divided into small areas. Therefore, an accurate directional feature vector of the contour can be obtained by matching the character stroke shifted from the direction of the dividing line to the dividing line. In addition, the image of the character stroke to be rotated is assumed to be a character stroke in each of upward and downward directions. Therefore, it is possible to accurately extract the direction feature vector of the oblique outline.

[0009]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of an apparatus to which the present invention is applied. Referring to FIG.
1, a touch panel 2, an LCD 3 formed on the lower surface of the touch panel, a memory 4, and a dictionary memory 5. Here, when the touch panel 2 is moved while being pressed by an input pen (not shown) between points A and B on the touch panel 2, for example, the CPU 1 sets the movement locus of the input pen during this time as a plurality of coordinate values via the touch panel 2. input. Then, the line segment from the point A to the point B is stored in the memory 4 and the line segment is displayed on the corresponding portion of the LCD 3. In this manner, handwritten characters based on pen input can be displayed on the LCD 3.

In addition, the data stored in the memory 4 and L
The CPU 1 performs the following processes on the handwritten characters displayed on the CD 3, and compares the processing result with each of the character types stored in the dictionary memory 5. Recognize whether it corresponds to. That is, in this case, the CPU 1 first performs preprocessing to remove noise and the like from the input characters (handwritten characters), which are the characters to be identified, to make the characters smooth. Thereafter, a contour image as shown in FIG. 2B is obtained from the original image shown in FIG. 2A, and the contour image is divided into 64 × 64 pixels.

Then, with respect to all the pixels (black pixels) constituting the outline, the direction component of each pixel is detected. That is, an arbitrary black pixel constituting the contour is focused on, and a directional component of the pixel of interest is detected from a connection pattern between the pixel of interest and black pixels before and after the pixel of interest.
FIG. 3 is a diagram showing a detection state of such a directional component of a black pixel.

Here, in FIG. 3, a pixel a in a hatched portion
Ｃc indicate black pixels constituting the contour line, and pixels other than the hatched portions are assumed to be white pixels. If the pixel of interest is b and the black pixels before and after it are a and c, respectively, the outline is
The pixel of interest b enters the target pixel b horizontally from the black pixel a, and further reaches the black pixel c in the upward right direction through the pixel of interest b. Therefore, the target pixel b has two directional components of a horizontal component and a right upward component. It can be seen that Then, these two directional components are detected as the directional components of the pixel of interest b. Note that, in addition to the above-described two directional components, there are a vertical component and a rightward downward component as the directional components of the black pixels forming the contour line.
Two directional components of the one directional component are allocated.

When extracting the directional pattern of the contour, the CPU 1 sequentially detects the directional component for each black pixel on all the black pixels on the contour and based on the detected directional component of each black pixel. Is extracted. In this way, from the original figure (contour image) shown in FIG. 4A, horizontal, vertical, downward-sloping and upward-sloping directional patterns shown in FIG. 4B are generated.

On the other hand, a line interval is obtained from the original figure shown in FIG. 4A to obtain line density information as shown in FIG. 4C. Here, the original figure is divided into 8 × 8 regions, and a non-linear normalization process for determining the position of the dividing line is performed so that the line density of each region at this time is the same as the line density information. FIG. 4D shows the result of the non-linear normalization process. Then, a nonlinear mapping is performed on each direction pattern shown in FIG. 4B based on the nonlinear normalized information indicating the position of the dividing line, and a processing result for each direction pattern as shown in FIG. obtain.

After the completion of the non-linear normalization process by the CPU 1, the CPU 1 divides each of the directional patterns shown in FIG. The number of black pixels is counted, and a count value for each small area is generated as a directional feature vector of the pattern. FIG. 5 is a diagram showing an example of such equal division.
FIG. 5B shows a horizontal character stroke, and FIG.
FIG. 5D and FIG.
(E) shows each character stroke in the rightward and downward directions.

Here, when an attempt is made to obtain a feature amount (ie, directional feature vector amount) from the number of black pixels in each small area, FIG.
Although accurate information can be extracted for horizontal and vertical character strokes as shown in (b) and (c), FIG.
For character strokes in oblique directions as shown in (d) and (e), continuous stroke information cannot be obtained because they do not match the direction of the region dividing line, and thus accurate directional feature vectors cannot be obtained. There is a problem. For this reason,
For each of the character strokes in the oblique direction pattern shown in FIGS. 5D and 5E, the image is converted into the image shown in FIGS. 5H and 5I by rotating the image by 45 ° counterclockwise. Then, a direction feature vector can be extracted from this image.

The rotation conversion of such an image is performed based on simultaneous equations consisting of the following equations (2) and (3). That is, the coordinates of the converted character stroke are x ′,
If y ′, the x coordinate of the character stroke before conversion is: x = (x′−y ′) / 2 ^1/2 (2) The y coordinate of the character stroke before conversion is y = (x ′ + y) ') / 2 ^1/2 (3).

Therefore, the coordinate position (x ', y') after rotation
And the original (before rotation) character stroke coordinate position (x, y)
Is obtained from Equations (2) and (3), and the value of the coordinate position of the original character stroke is moved to the rotated coordinate position, whereby the image (graphic) can be rotated by 45 °. . Thus, for the character stroke rotated by 45 °, an 8 × character stroke as shown in FIGS.
Eight divisions are made equally. Thereafter, the number of black pixels in each of the equally divided square small areas is counted, and the directional feature vector for the character stroke in the downward right direction and the direction for the character stroke in the upward right direction are calculated based on the count value of each small area. Get the feature vector.

In this way, the direction feature vector can be obtained. Thereafter, for the obtained direction feature vector, a plurality of candidate characters having a vector with a short distance among vectors of the dictionary in the dictionary memory 5 are extracted in order from the one with the shortest distance, and the extracted candidate character and the input character are extracted. By comparing in more detail, final character identification is performed. Table 1
Fig. 6 shows the results of character recognition experiments based on the directional feature vector extraction method of the present invention and the results of character recognition by a conventional method without rotating an image, as recognition rates (%). The first-order classification rate, third-order classification rate,... In the figure indicate the order when the sample of the recognized character is selected as a candidate character from the dictionary.

[0020]

[Table 1]

Here, a comparison between the results of the recognition experiment by the directional feature vector extraction method of the present invention and the results of the recognition experiment by the conventional method shows that the character when the directional feature vector extraction method of the present invention is used at any classification rate. It can be seen that the recognition rate is higher than the conventional method. As described above, when the directional feature vector is extracted, the image in which the stroke in the oblique direction is the main component is rotated by 45 ° to be converted into the image in the horizontal or vertical direction. Therefore, it is possible to extract continuous stroke information as well as horizontal or vertical character strokes, and thus it is possible to extract an accurate directional feature vector of an image.

In this embodiment, a character stroke having upward and downward directional components is rotated by 45 °. However, by setting the rotation angle freely, horizontal or vertical characters can be rotated. It can also be applied to strokes. Generally, when writing a character, a character rising to the right is often written. In such a case, the character stroke in the horizontal direction is rotated by a predetermined angle so that the character rising to the right is made closer to the horizontal direction, and a more accurate character stroke can be obtained. Therefore, an accurate direction feature vector is extracted, and accurate character recognition can be performed. In addition,
The same applies to the character stroke of the vertical component.
In this embodiment, an example of handwritten characters has been described. However, any character data may be used as long as the data is bitmap data, such as character data read by OCR.

[0023]

As described above, according to the present invention, the outline is extracted from the black pixels constituting the outline image of the character to be identified as horizontal, vertical, upward-sloping and downward-sloping patterns. The contour image is normalized, and the above-described directional patterns are processed based on the normalized information to generate respective character strokes.The generated character strokes in the respective directions are divided into small areas by horizontal and vertical dividing lines. When the number of black pixels in each divided small area is equally counted and extracted as a directional feature vector, the generated image of the character stroke is rotated so as to match the direction of the dividing line, and rotated. Subsequent character strokes are equally divided into small areas, so that the images of character strokes that deviate from the direction of the dividing line match Click information is obtained, this result, since the correct orientation feature vector of the contour line is obtained, it is possible to perform accurate character recognition. In addition, since the image of the character stroke to be rotated is a character stroke in each of the upward and downward directions, the direction feature vector of the oblique outline can be accurately extracted.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an apparatus to which the present invention is applied.

FIG. 2 is a diagram showing an original image to be recognized by the present apparatus and a contour line of the original image.

FIG. 3 is a diagram illustrating a detection state of a directional component of a pixel in the present apparatus.

FIG. 4 is a diagram showing a state of extraction of a direction pattern of a contour line and a state of nonlinear normalization processing.

FIG. 5 is a diagram showing a rotation state of a character stroke when a direction feature vector is obtained.

FIG. 6 is a diagram illustrating an example of an oblique contour line from which an accurate directional feature vector cannot be obtained.

FIG. 7 is a recognition processing flow showing a conventional character recognition process.

[Explanation of symbols]

1 ... CPU, 2 ... Touch panel, 3 ... LCD, 4 ... Memory, 5 ... Dictionary memory

Claims (2)

[Claims] 1. A contour line is extracted as horizontal, vertical, upward and downward directional patterns from black pixels constituting a contour image of a character to be identified, and the contour image is normalized to obtain the normalized information. The above-described direction patterns are processed based on the generated character strokes, and the generated character strokes in each direction are equally divided into small areas by horizontal and vertical dividing lines. In the directional feature vector extraction method of counting the number of pixels for each pixel and extracting it as a directional feature vector, the generated character stroke image is rotated so as to match the direction of the dividing line, and the rotated character stroke is rotated. A directional feature vector extraction method, comprising equally dividing the small area. 2. The directional feature vector extraction method according to claim 1, wherein the image of the character stroke to be rotated is a character stroke in each of upward and downward directions.